• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1###############################################################################
2# Server process to keep track of unlinked resources (like shared memory
3# segments, semaphores etc.) and clean them.
4#
5# On Unix we run a server process which keeps track of unlinked
6# resources. The server ignores SIGINT and SIGTERM and reads from a
7# pipe.  Every other process of the program has a copy of the writable
8# end of the pipe, so we get EOF when all other processes have exited.
9# Then the server process unlinks any remaining resource names.
10#
11# This is important because there may be system limits for such resources: for
12# instance, the system only supports a limited number of named semaphores, and
13# shared-memory segments live in the RAM. If a python process leaks such a
14# resource, this resource will not be removed till the next reboot.  Without
15# this resource tracker process, "killall python" would probably leave unlinked
16# resources.
17
18import os
19import signal
20import sys
21import threading
22import warnings
23
24from . import spawn
25from . import util
26
27__all__ = ['ensure_running', 'register', 'unregister']
28
29_HAVE_SIGMASK = hasattr(signal, 'pthread_sigmask')
30_IGNORED_SIGNALS = (signal.SIGINT, signal.SIGTERM)
31
32def cleanup_noop(name):
33    raise RuntimeError('noop should never be registered or cleaned up')
34
35_CLEANUP_FUNCS = {
36    'noop': cleanup_noop,
37    'dummy': lambda name: None,  # Dummy resource used in tests
38}
39
40if os.name == 'posix':
41    import _multiprocessing
42    import _posixshmem
43
44    # Use sem_unlink() to clean up named semaphores.
45    #
46    # sem_unlink() may be missing if the Python build process detected the
47    # absence of POSIX named semaphores. In that case, no named semaphores were
48    # ever opened, so no cleanup would be necessary.
49    if hasattr(_multiprocessing, 'sem_unlink'):
50        _CLEANUP_FUNCS.update({
51            'semaphore': _multiprocessing.sem_unlink,
52        })
53    _CLEANUP_FUNCS.update({
54        'shared_memory': _posixshmem.shm_unlink,
55    })
56
57
58class ReentrantCallError(RuntimeError):
59    pass
60
61
62class ResourceTracker(object):
63
64    def __init__(self):
65        self._lock = threading.RLock()
66        self._fd = None
67        self._pid = None
68        self._exitcode = None
69
70    def _reentrant_call_error(self):
71        # gh-109629: this happens if an explicit call to the ResourceTracker
72        # gets interrupted by a garbage collection, invoking a finalizer (*)
73        # that itself calls back into ResourceTracker.
74        #   (*) for example the SemLock finalizer
75        raise ReentrantCallError(
76            "Reentrant call into the multiprocessing resource tracker")
77
78    def _stop(self):
79        with self._lock:
80            # This should not happen (_stop() isn't called by a finalizer)
81            # but we check for it anyway.
82            if self._lock._recursion_count() > 1:
83                return self._reentrant_call_error()
84            if self._fd is None:
85                # not running
86                return
87
88            # closing the "alive" file descriptor stops main()
89            os.close(self._fd)
90            self._fd = None
91
92            _, status = os.waitpid(self._pid, 0)
93
94            self._pid = None
95
96            try:
97                self._exitcode = os.waitstatus_to_exitcode(status)
98            except ValueError:
99                # os.waitstatus_to_exitcode may raise an exception for invalid values
100                self._exitcode = None
101
102    def getfd(self):
103        self.ensure_running()
104        return self._fd
105
106    def ensure_running(self):
107        '''Make sure that resource tracker process is running.
108
109        This can be run from any process.  Usually a child process will use
110        the resource created by its parent.'''
111        with self._lock:
112            if self._lock._recursion_count() > 1:
113                # The code below is certainly not reentrant-safe, so bail out
114                return self._reentrant_call_error()
115            if self._fd is not None:
116                # resource tracker was launched before, is it still running?
117                if self._check_alive():
118                    # => still alive
119                    return
120                # => dead, launch it again
121                os.close(self._fd)
122
123                # Clean-up to avoid dangling processes.
124                try:
125                    # _pid can be None if this process is a child from another
126                    # python process, which has started the resource_tracker.
127                    if self._pid is not None:
128                        os.waitpid(self._pid, 0)
129                except ChildProcessError:
130                    # The resource_tracker has already been terminated.
131                    pass
132                self._fd = None
133                self._pid = None
134                self._exitcode = None
135
136                warnings.warn('resource_tracker: process died unexpectedly, '
137                              'relaunching.  Some resources might leak.')
138
139            fds_to_pass = []
140            try:
141                fds_to_pass.append(sys.stderr.fileno())
142            except Exception:
143                pass
144            cmd = 'from multiprocessing.resource_tracker import main;main(%d)'
145            r, w = os.pipe()
146            try:
147                fds_to_pass.append(r)
148                # process will out live us, so no need to wait on pid
149                exe = spawn.get_executable()
150                args = [exe] + util._args_from_interpreter_flags()
151                args += ['-c', cmd % r]
152                # bpo-33613: Register a signal mask that will block the signals.
153                # This signal mask will be inherited by the child that is going
154                # to be spawned and will protect the child from a race condition
155                # that can make the child die before it registers signal handlers
156                # for SIGINT and SIGTERM. The mask is unregistered after spawning
157                # the child.
158                try:
159                    if _HAVE_SIGMASK:
160                        signal.pthread_sigmask(signal.SIG_BLOCK, _IGNORED_SIGNALS)
161                    pid = util.spawnv_passfds(exe, args, fds_to_pass)
162                finally:
163                    if _HAVE_SIGMASK:
164                        signal.pthread_sigmask(signal.SIG_UNBLOCK, _IGNORED_SIGNALS)
165            except:
166                os.close(w)
167                raise
168            else:
169                self._fd = w
170                self._pid = pid
171            finally:
172                os.close(r)
173
174    def _check_alive(self):
175        '''Check that the pipe has not been closed by sending a probe.'''
176        try:
177            # We cannot use send here as it calls ensure_running, creating
178            # a cycle.
179            os.write(self._fd, b'PROBE:0:noop\n')
180        except OSError:
181            return False
182        else:
183            return True
184
185    def register(self, name, rtype):
186        '''Register name of resource with resource tracker.'''
187        self._send('REGISTER', name, rtype)
188
189    def unregister(self, name, rtype):
190        '''Unregister name of resource with resource tracker.'''
191        self._send('UNREGISTER', name, rtype)
192
193    def _send(self, cmd, name, rtype):
194        try:
195            self.ensure_running()
196        except ReentrantCallError:
197            # The code below might or might not work, depending on whether
198            # the resource tracker was already running and still alive.
199            # Better warn the user.
200            # (XXX is warnings.warn itself reentrant-safe? :-)
201            warnings.warn(
202                f"ResourceTracker called reentrantly for resource cleanup, "
203                f"which is unsupported. "
204                f"The {rtype} object {name!r} might leak.")
205        msg = '{0}:{1}:{2}\n'.format(cmd, name, rtype).encode('ascii')
206        if len(msg) > 512:
207            # posix guarantees that writes to a pipe of less than PIPE_BUF
208            # bytes are atomic, and that PIPE_BUF >= 512
209            raise ValueError('msg too long')
210        nbytes = os.write(self._fd, msg)
211        assert nbytes == len(msg), "nbytes {0:n} but len(msg) {1:n}".format(
212            nbytes, len(msg))
213
214
215_resource_tracker = ResourceTracker()
216ensure_running = _resource_tracker.ensure_running
217register = _resource_tracker.register
218unregister = _resource_tracker.unregister
219getfd = _resource_tracker.getfd
220
221
222def main(fd):
223    '''Run resource tracker.'''
224    # protect the process from ^C and "killall python" etc
225    signal.signal(signal.SIGINT, signal.SIG_IGN)
226    signal.signal(signal.SIGTERM, signal.SIG_IGN)
227    if _HAVE_SIGMASK:
228        signal.pthread_sigmask(signal.SIG_UNBLOCK, _IGNORED_SIGNALS)
229
230    for f in (sys.stdin, sys.stdout):
231        try:
232            f.close()
233        except Exception:
234            pass
235
236    cache = {rtype: set() for rtype in _CLEANUP_FUNCS.keys()}
237    exit_code = 0
238
239    try:
240        # keep track of registered/unregistered resources
241        with open(fd, 'rb') as f:
242            for line in f:
243                try:
244                    cmd, name, rtype = line.strip().decode('ascii').split(':')
245                    cleanup_func = _CLEANUP_FUNCS.get(rtype, None)
246                    if cleanup_func is None:
247                        raise ValueError(
248                            f'Cannot register {name} for automatic cleanup: '
249                            f'unknown resource type {rtype}')
250
251                    if cmd == 'REGISTER':
252                        cache[rtype].add(name)
253                    elif cmd == 'UNREGISTER':
254                        cache[rtype].remove(name)
255                    elif cmd == 'PROBE':
256                        pass
257                    else:
258                        raise RuntimeError('unrecognized command %r' % cmd)
259                except Exception:
260                    exit_code = 3
261                    try:
262                        sys.excepthook(*sys.exc_info())
263                    except:
264                        pass
265    finally:
266        # all processes have terminated; cleanup any remaining resources
267        for rtype, rtype_cache in cache.items():
268            if rtype_cache:
269                try:
270                    exit_code = 1
271                    if rtype == 'dummy':
272                        # The test 'dummy' resource is expected to leak.
273                        # We skip the warning (and *only* the warning) for it.
274                        pass
275                    else:
276                        warnings.warn(
277                            f'resource_tracker: There appear to be '
278                            f'{len(rtype_cache)} leaked {rtype} objects to '
279                            f'clean up at shutdown: {rtype_cache}'
280                        )
281                except Exception:
282                    pass
283            for name in rtype_cache:
284                # For some reason the process which created and registered this
285                # resource has failed to unregister it. Presumably it has
286                # died.  We therefore unlink it.
287                try:
288                    try:
289                        _CLEANUP_FUNCS[rtype](name)
290                    except Exception as e:
291                        exit_code = 2
292                        warnings.warn('resource_tracker: %r: %s' % (name, e))
293                finally:
294                    pass
295
296        sys.exit(exit_code)
297